OpenAI's O-series models, DeepSeek R1/R1 zero, and xAI's recently launched Grok 3 are all equipped with reasoning capabilities, but according to KOLs and related papers on the Internet, we can still speculate on their core technologies and design concepts:

• Adding Reinforcement Learning to the original Pretraining process in LLM.
• In the final Inference Model (Inference Model) also added Reinforcement Learning (Reinforcement Learning) and Monte Carlo Tree Search (MCTS), etc., mainly to enhance or improve the test computing resources (Test-Time Compute).

In terms of end-inference users, the core mechanism of O3 is modeled in the